Stop Building Workflows, Start Describing Them: Comfy Agent Is Live
For years, ComfyUI has given creators unprecedented control: a node-based canvas where you wire together image generation, video synthesis, upscaling, masking, and everything else into a single graph. The problem? That flexibility requires knowledge. You need to know which nodes exist, how to connect them, what each parameter does. Most people who want to make something complex end up either copying workflows or giving up.
This week, Comfy Org launched Comfy Agent to flatten that curve. It's an AI teammate that works directly on your canvas, building workflows in real time while you stay in edit mode. You describe what you want—"take this product photo and make 20 on-brand ad variants in 4 aspect ratios"—and it plans the pipeline, drops nodes, wires them together, runs the graph, and iterates when something breaks.
What's Different About How It Works
Comfy Agent runs on frontier models by Anthropic, but the real trick is that it doesn't hand you JSON to import. It plans the workflow, drops the nodes onto the canvas, wires them together, runs the graph and iterates when something breaks, working on the canvas you already have open rather than handing back a JSON file to import, so the graph stays visible and editable while it builds.
That means you can drag an asset from your library, mention a node by name, or move things around while the agent is still building. The agent works directly on your canvas while you edit the workflow too, designed this way so both of you won't have to wait for or interrupt each other. What the model does not get on its own, the agent adds: the full context of your workflow, your errors and ComfyUI itself.
It is open to everyone on Comfy Cloud from day one, and a version for Comfy Desktop is promised in a few weeks. Free initial tokens are now available for everyone to try several Comfy Agent chats, and after that, Comfy Agent charges from the same Comfy Credits pool you already use for generations.
What People Are Actually Making With It
Three creators worked with Comfy Agent ahead of the October 1 launch. An animator working on the story of an original character brought characters and imagined worlds into one creative workflow, with the agent helping to shape the look, explore scenes, and compare generation results—one artist, 100 hours, working with Comfy Agent.
A VFX use case by Doug Hogan isolates, masks, and refines—taking a shot from image to compositing passes, isolating the subject, generating masks, and working with depth and normals in one workflow.
On the product side, the example Comfy Org highlights is straightforward: batch product photography. Ask it for something real: "Take this product photo and make 20 on-brand ad variants in 4 aspect ratios." Create 3 product scenes from the base composition and bottle references—white, gold, and purple keyframes generated as product scenes, then use the motion reference and keyframe to render a video ad.
What's Actually Good Here
The real strength is specificity. Comfy Agent isn't a generic chatbot that hallucinates nodes. It plans, builds, runs and debugs ComfyUI workflows on the canvas itself, with reusable skills and up to 5 parallel chats. You can save "skills"—reusable workflow patterns you write once and call again—which means the agent learns what your studio does.
The real-time collaboration part matters too. If you're used to describing something and getting back a broken workflow, this is different. You can see the graph form and redirect it mid-build, which cuts out the "I'll just copy someone else's workflow and hack it" moment.
What Isn't There Yet
Local support (Comfy Desktop) is still coming. Once Comfy Agent reaches Comfy Desktop, it will work with your own GPU, models, and custom nodes, and walk you through setup, and will also build custom editing tools directly on the canvas for you, such as a one-off masking or compositing tool you didn't have to install. Until then, you're on Comfy Cloud, which means network requests and no offline builds.
Local LLM support is on the roadmap, which some builders are waiting for—not everyone wants to stream requests to Anthropic's models.
How to Try It
Log in at cloud.comfy.org, open the "Agent" panel to claim free initial chats, and ask it for something real: "Take this product photo and make 20 on-brand ad variants in 4 aspect ratios." The free tokens give you a few conversations to test the concept. After that, it runs on your existing Comfy credits.